Grey-Box Digital Twin Scaling for Bio-Chemical Process Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Digital twin technology in chemical and biochemical production plants faces challenges due to high expertise and time requirements for modeling complex physio-chemical phenomena, especially in the specialty chemical industry where material properties and reaction kinetics are rarely available, and scaling machine learning-based models from lab to production is not effectively solved.

Innovation Solution

A method using grey-box models that combine data-based machine learning with physical models, trained with production history data, and scaled using laboratory equipment to generate representative data, allowing for efficient optimization and scaling of production units, including the integration of artificial neural networks and specific software frameworks like PyTorch and CasADi for modeling and control strategies.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If fully data-based machine learning models are used, then flexibility and adaptability are improved, but reliability and scalability to production scale deteriorate

Engineering Contradiction:
Improvemodel flexibilityVSAvoidmodel reliability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent combines physics-based models (providing reliability and scalability) with machine learning models (providing flexibility and adaptability) into a hybrid grey-box model. The physics-based component ensures fundamental correctness while the ML component captures complex nonlinear behaviors, resolving the contradiction between reliability and adaptability.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The grey-box model acts as a composite modeling approach, integrating two different modeling paradigms (physics-based and data-based) into a unified framework. This composite structure leverages the strengths of both approaches while mitigating their individual weaknesses, enabling both reliability and flexibility.

Inventive Principle:
Principle #40Composite materials

2Reliability

If rigorous physical models are developed for novel processes, then model reliability is improved, but development time increases

Engineering Contradiction:
Improvemodel reliabilityVSAvoidmodel development time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

Instead of developing complete first-principles physical models for all process aspects, the patent applies physics-based models only to critical components (mass and energy balances) while using machine learning for other complex phenomena. This partial application of physics-based modeling reduces development time while maintaining sufficient reliability for scale-up.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent uses available production history data and literature information to pre-configure the physics-based model structure and parameters before detailed development. This preliminary action reduces the time required for rigorous model development by starting from a pre-established framework.

Inventive Principle:
Principle #10Preliminary action

3Quantity of substance

If production history data is used for model training, then data availability is improved, but data representativeness deteriorates due to recipe constraints

Engineering Contradiction:
Improvedata availabilityVSAvoiddata representativeness
Core Design Contradiction:
Quantity of substanceVSManufacturing precision

Solution Approach 1:

The patent introduces a data transformation and augmentation process as an intermediary between production history data and model training. This includes techniques like noise injection, operating point perturbation, and synthetic data generation that enhance the representativeness of production data while preserving its availability.

Inventive Principle:
Principle #24Intermediary (Mediator)

4Manufacturing precision

If lab-scale equipment is used to generate representative data, then data representativeness is improved, but scaling difficulty increases

Engineering Contradiction:
Improvedata representativenessVSAvoidscaling complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent systematically varies key operating parameters (temperature, pressure, flow rates, concentrations) during lab-scale experiments to generate diverse training data. This parameter exploration approach ensures data representativeness while the data is collected at lab-scale, avoiding the need to scale up equipment.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

Instead of scaling up the physical equipment dimension, the patent explores the parameter space dimension by conducting experiments across a wide range of operating conditions at lab-scale. This dimensional shift allows generating representative data without scaling complexity.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4270120A1Embedded model-based digital twin workflow for the accelerated optimization of bio-/ chemical processes
Publication Date: 2023.11.01 MERCK PATENT GMBH
  • EP4270120A1 patent drawingFigure 1
  • EP4270120A1 patent drawingFigure 2
  • EP4270120A1 patent drawingFigure 3

AI summary

A method to improve bio-/chemical and/or pharmaceutical production units in a production plant controlled by at least one computer the following steps of Determining a grey-box model configuration wherein a data-based machine learning model component describes a function which is independent of the production plant equipment scale and wherein the data-based machine learning model component is embedded into a system of ordinary differential equations or differential-algebraic equations which express the component mass and energy balances of the modeled process; Setting up the grey-box model via a software running on the at least one computer including the machine learning model component and a physical model component with the determined configuration and the dimensions of the required production scale of the production unit, and training its machine learning model component using available production history data; Setting up a downscaled laboratory version of the production unit operation, and scale the set up grey-box model to match it via its physical parameters; Perform experiments to generate representative data for extending the validity of the machine learning component of the grey-box model therefore improving it; Identifying at least one new operating window for the production unit using the improved grey-box model and numerical optimization techniques and validating the model-based control strategy with it; and Scaling up the grey-box model and the control strategy to the required production scale and performing validation runs, comprising. The core of the invention is a workflow to accelerate the optimization of existing production units and the scale-up of new processes by using a specific configuration for grey-box models that enables straightforward scaling, and representative data generation with laboratory equipment.